• 목록
• 게시판 리스트 옵션
  • 검색

In a semiconductor manufacturing process, it is important to detect and classify defect patterns in Wafer Bin Maps (WBMs) and identify the root cause of these defects for tight quality control. Recently, various deep learning methods have been applied, but these methods suffer from poor classification, limited data labelling, and the inability to detect and learn new defect patterns. Moreover, the methods prioritize improving the accuracy and speed of classification models over detecting and classifying unknown defect patterns. Against this background, we developed an abnormal pattern detector based on a One-Class Support Vector Machine (SVM) that classifies whether defect patterns are known or unknown. For the known patterns, we used transfer learning based on a ResNet50 classifier pre-trained with ImageNet1K data to further classify the defect pattern in the WBM. For the unknown patterns, clustering is performed using Density-Based Spatial Clustering Application with Noise (DBSCAN) to assign new labels, and the classifier is updated through active learning. This enables the detection of unknown patterns and effectively updates the abnormal detector and pre-trained classifier even during the use of the classifier. Experiment results from the WM-811K dataset verify that the proposed method can detect unknown patterns while maintaining excellent classification performance for known patterns. Moreover, it can continuously maintain the high classification performance of the detector and pre-trained classifier through the active learning. Also, applicability in real semiconductor manufacturing environments was demonstrated using real industrial data with an unknown pattern (“Eye Defect Pattern”), not included in the WM-811K dataset.